perf: bypass Celery for web chat — stream LLM directly from WebSocket

Eliminates 5-10s of overhead by calling the LLM pool's streaming
endpoint directly from the WebSocket handler instead of going through
Celery queue → worker → asyncio.run() → Redis pub-sub → WebSocket.

New flow: WebSocket → agent lookup → memory → LLM stream → WebSocket
Old flow: WebSocket → Celery → worker → DB → memory → LLM → Redis → WebSocket

Memory still saved (Redis sliding window + fire-and-forget embedding).
Slack/WhatsApp still use Celery (async webhook pattern).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

packages/gateway/gateway/channels/web.py

@@ -42,11 +42,15 @@ import redis.asyncio as aioredis
 from fastapi import APIRouter, WebSocket, WebSocketDisconnect
 from sqlalchemy import select, text
 
-from orchestrator.tasks import handle_message
+from orchestrator.agents.builder import build_messages_with_memory, build_system_prompt
+from orchestrator.agents.runner import run_agent_streaming
+from orchestrator.memory.short_term import get_recent_messages, append_message
+from orchestrator.tasks import handle_message, embed_and_store
 from shared.config import settings
 from shared.db import async_session_factory, engine
 from shared.models.chat import WebConversation, WebConversationMessage
 from shared.models.message import ChannelType, KonstructMessage, MessageContent, SenderInfo
+from shared.models.tenant import Agent
 from shared.redis_keys import webchat_response_key
 from shared.rls import configure_rls_hook, current_tenant_id
 
@@ -225,7 +229,12 @@ async def _handle_websocket_connection(
             current_tenant_id.reset(rls_token)
 
         # -------------------------------------------------------------------
-        # c. Normalize and dispatch to Celery pipeline
+        # c. Build KonstructMessage and stream LLM response DIRECTLY
+        #
+        # Bypasses Celery entirely for web chat — calls the LLM pool's
+        # streaming endpoint from the WebSocket handler. This eliminates
+        # ~5-10s of Celery queue + Redis pub-sub round-trip overhead.
+        # Slack/WhatsApp still use Celery (async webhook pattern).
         # -------------------------------------------------------------------
         event = {
             "text": text_content,
@@ -237,67 +246,92 @@ async def _handle_websocket_connection(
         }
         normalized_msg = normalize_web_event(event)
 
-        extras = {
+        # Load agent for this tenant
-            "conversation_id": saved_conversation_id,
+        agent: Agent | None = None
-            "portal_user_id": user_id_str,
+        rls_token3 = current_tenant_id.set(tenant_uuid)
-        }
-        task_payload = normalized_msg.model_dump(mode="json") | extras
-        handle_message.delay(task_payload)
-
-        # -------------------------------------------------------------------
-        # d. Subscribe to Redis pub-sub and forward streaming chunks to client
-        #
-        # The orchestrator publishes two message types to the response channel:
-        #   {"type": "chunk", "text": "<token>"}  — zero or more times (streaming)
-        #   {"type": "done",  "text": "<full>", "conversation_id": "..."}  — final marker
-        #
-        # We forward each "chunk" immediately to the browser so text appears
-        # word-by-word. On "done" we save the full response to the DB.
-        # -------------------------------------------------------------------
-        response_channel = webchat_response_key(tenant_id_str, saved_conversation_id)
-        subscribe_redis = aioredis.from_url(settings.redis_url)
-        response_text: str = ""
         try:
-            pubsub = subscribe_redis.pubsub()
+            async with async_session_factory() as session:
-            await pubsub.subscribe(response_channel)
+                from sqlalchemy import select as sa_select
-
+                agent_stmt = sa_select(Agent).where(
-            deadline = asyncio.get_event_loop().time() + _RESPONSE_TIMEOUT_SECONDS
+                    Agent.tenant_id == tenant_uuid,
-
+                    Agent.is_active == True,
-            while asyncio.get_event_loop().time() < deadline:
+                ).limit(1)
-                message = await pubsub.get_message(ignore_subscribe_messages=True, timeout=1.0)
+                agent_result = await session.execute(agent_stmt)
-                if message and message.get("type") == "message":
+                agent = agent_result.scalar_one_or_none()
-                    try:
-                        payload = json.loads(message["data"])
-                    except (json.JSONDecodeError, KeyError):
-                        await asyncio.sleep(0.01)
-                        continue
-
-                    msg_type = payload.get("type")
-
-                    if msg_type == "chunk":
-                        # Forward token immediately — do not break the loop
-                        token = payload.get("text", "")
-                        if token:
-                            try:
-                                await websocket.send_json({
-                                    "type": "chunk",
-                                    "text": token,
-                                })
-                            except Exception:
-                                # Client disconnected mid-stream — exit cleanly
-                                break
-
-                    elif msg_type == "done":
-                        # Final marker — full text for DB persistence
-                        response_text = payload.get("text", "")
-                        break
-
-                else:
-                    await asyncio.sleep(0.05)
-
-            await pubsub.unsubscribe(response_channel)
         finally:
-            await subscribe_redis.aclose()
+            current_tenant_id.reset(rls_token3)
+
+        if agent is None:
+            await websocket.send_json({
+                "type": "done",
+                "text": "No active AI employee is configured for this workspace.",
+                "conversation_id": saved_conversation_id,
+            })
+            continue
+
+        # Build memory-enriched messages (Redis sliding window only — fast)
+        redis_mem = aioredis.from_url(settings.redis_url)
+        try:
+            recent_messages = await get_recent_messages(
+                redis_mem, tenant_id_str, str(agent.id), user_id_str
+            )
+        finally:
+            await redis_mem.aclose()
+
+        enriched_messages = build_messages_with_memory(
+            agent=agent,
+            current_message=text_content,
+            recent_messages=recent_messages,
+            relevant_context=[],
+            channel="web",
+        )
+
+        # Stream LLM response directly to WebSocket — no Celery, no pub-sub
+        response_text = ""
+        try:
+            async for token in run_agent_streaming(
+                msg=normalized_msg,
+                agent=agent,
+                messages=enriched_messages,
+            ):
+                response_text += token
+                try:
+                    await websocket.send_json({"type": "chunk", "text": token})
+                except Exception:
+                    break  # Client disconnected
+        except Exception:
+            logger.exception("Direct streaming failed for conversation=%s", saved_conversation_id)
+            if not response_text:
+                response_text = "I encountered an error processing your message. Please try again."
+
+        # Save to Redis sliding window (fire-and-forget, non-blocking)
+        redis_mem2 = aioredis.from_url(settings.redis_url)
+        try:
+            await append_message(redis_mem2, tenant_id_str, str(agent.id), user_id_str, "user", text_content)
+            if response_text:
+                await append_message(redis_mem2, tenant_id_str, str(agent.id), user_id_str, "assistant", response_text)
+        finally:
+            await redis_mem2.aclose()
+
+        # Fire-and-forget embedding for long-term memory
+        try:
+            embed_and_store.delay({
+                "tenant_id": tenant_id_str,
+                "agent_id": str(agent.id),
+                "user_id": user_id_str,
+                "role": "user",
+                "content": text_content,
+            })
+            if response_text:
+                embed_and_store.delay({
+                    "tenant_id": tenant_id_str,
+                    "agent_id": str(agent.id),
+                    "user_id": user_id_str,
+                    "role": "assistant",
+                    "content": response_text,
+                })
+        except Exception:
+            pass  # Non-fatal — memory will rebuild over time
 
         # -------------------------------------------------------------------
         # e. Save assistant message and send final "done" to client